AI and the Empathy Paradox

There is a moment in therapy that most of us recognize, even if we don’t often name it directly.

A client pauses, looks at us, and says something like: “I know you’re trying to be empathic… but I don’t feel it.” Have you ever had this experience?

It can land as a rupture, a challenge, or even a kind of disappointment. And yet, these moments reveal, in a way that is difficult to ignore, that empathy in psychotherapy is not reducible to accuracy. It is not simply a matter of saying the right words, reflecting the correct emotion, or demonstrating that we understand.

In a recent conversation between Bruce Wampold and Anat Perry, a social cognitive neuroscientist who studies empathy, a striking tension emerged. Artificial intelligence (AI) systems are already capable of generating what many would consider excellent empathic responses. They can reflect, validate, and articulate emotional understanding with a level of fluency that, in some cases, rivals or even exceeds that of many human therapists. And yet, when people discover that these responses are generated by AI, something shifts. The same words, received moments earlier as helpful or even moving, can suddenly feel hollow, even objectionable.

This points to something more fundamental about how empathy functions in human relationships. Perry proposes a useful distinction between three components of empathy: cognitive empathy (the ability to understand another person’s emotional state), affective empathy (the capacity to share or resonate with that state), and motivational empathy (the desire to alleviate another’s suffering). AI systems, she suggests, can approximate aspects of cognitive empathy quite effectively. They can organize information, detect patterns, and generate responses that mirror what a person might be feeling. They can even simulate the language of care and concern. But simulation is not the same as experience, and at some level, people seem to register that difference. She has an elegant publication about the theme, and you can check it out here.

For clinicians, this may resonate with something we already encounter in practice. Two therapists can offer nearly identical reflections, yet one lands and the other does not. The difference is not always in the wording. It is often in something more difficult to specify, something about presence, about being affected, about the sense that there is another mind in the room that is not only observing but participating.

Dr. Bruce Wampold makes an important observation in this regard: therapy is not built on consistently getting empathy “right.” In fact, some of the most meaningful moments in therapy occur when empathy initially fails. When the therapist misunderstands, when the client feels unseen, and when that gap becomes something that can be explored and repaired. These moments of rupture and repair are not peripheral to the work; they are constitutive of it. They allow the relationship to become something dynamic, something negotiated, something real.

AI, by contrast, tends to bypass this process. It offers responses that are immediately coherent, attuned, and “correct.” But in doing so, it removes the possibility of misattunement, and with it, the opportunity for repair. What is lost is not just imperfection, but the relational process through which meaning is co-constructed.

Another way to approach this is through the lens of effort. One of the more subtle insights from Perry’s work is that empathy is not only about what is communicated, but about what is inferred. When we receive an empathic response from another person, we are not only processing the content of their words; we are also, often implicitly, registering that someone has taken the time to think about us, to feel with us, to try to understand. There is a sense that something has been given. AI-generated responses, even when highly articulate, do not carry the same signal of effort. And without that signal, something about the exchange can feel less meaningful.

At the same time, it would be overly simplistic to frame this as a binary between “real” human therapy and “artificial” AI interaction. The conversation also highlights domains in which AI may offer genuine value. When what a person needs is structure, guidance, or specific tools such as psychoeducation, behavioral strategies, or help organizing thoughts. AI systems may be well suited to provide this type of support, and this is not entirely new; forms of bibliotherapy and guided self-help have long demonstrated effectiveness for certain concerns. What AI adds is immediacy, personalization, and interactivity.

The more complex question arises when we consider how these functions intersect with relational needs. In practice, therapy is rarely only about tools or only about connection. Most clients move between these domains, often within the same session. They may seek concrete strategies for managing anxiety while also needing to feel understood in the context of their broader life experience, and the feeling of connection with another human being. The possibility of integrating AI into this process, perhaps as a supplement between sessions, or as a way to support skill acquisition, raises interesting possibilities. But it also requires careful attention to what is being supplemented, and what must remain grounded in human relationships.

There is also a broader social context that cannot be ignored. We are already living in what is often described as a loneliness epidemic. In this context, AI systems that are always available, consistently responsive, and nonjudgmental may be particularly appealing, especially for individuals who struggle to form or maintain relationships. And yet, these may be precisely the individuals for whom human connection is most critical. The risk is not necessarily that AI will replace therapy, but that it will become a substitute for the kinds of relational experiences that therapy, at its best, can help facilitate.

A further complication is beginning to emerge in everyday interactions. As AI tools become more integrated into communication, it may become increasingly difficult to discern whether expressions of care or empathy originate from a person or from a system. This ambiguity has the potential to erode trust in subtle and different ways. If we cannot reliably infer that another person has chosen their words, that they have reflected and responded in their own voice, the meaning of those words may begin to shift. Empathy, in this sense, risks losing some of its relational specificity.

For clinicians, these developments may feel both concerning and clarifying. On one hand, they raise important questions about the future of therapeutic practice and the broader ecology of human connection. On the other, they highlight what is distinctive about psychotherapy as a relational process. Therapy is not simply the delivery of well-formed empathic statements. It is an ongoing, imperfect, and often unpredictable encounter between two people. It involves misunderstanding and repair, effort and presence, and the gradual emergence of trust.

If there is a reassurance to be found here, it is that the core elements of therapeutic work are not easily replicated by systems that operate primarily through pattern recognition and response generation. At the same time, the emergence of these technologies invites us to be more precise about what we do, how we do it, and why it matters.

The question, then, may not be whether AI can do therapy, but rather how we, as clinicians, can engage with these tools in ways that support our work without displacing the relational conditions that make that work possible.

While much of the conversation around AI in psychotherapy centers on whether it can replace the therapist, a more immediately relevant development is emerging in how it may support the work around the session.

Platforms like CarePaths have introduced AI-assisted note generation – tools that can produce structured clinical documentation based on session content, which the clinician can then review and finalize. These systems are designed to meet documentation and compliance requirements without the same time commitment, offering flexible formats such as SOAP or DAP notes, and integrating into both teletherapy and in-person workflows.

At first glance, this appears to be a way to reduce time spent on documentation, streamline workflows, and address a well-known source of clinician burden. But viewed through a clinical lens, the implications are more nuanced.

One of the central concerns in any use of AI in clinical contexts is privacy. In CarePaths, these tools are built with attention to confidentiality: session recordings are not stored, and the data is not used to train external models. This distinction matters. The therapeutic space depends on a shared understanding that what is spoken remains contained, and any technological extension of that space must preserve this boundary.

At the same time, documentation itself is not neutral. It shapes how we attend in session, what we retain afterward, and how we metabolize clinical encounters. Many clinicians are familiar with the subtle division of attention that occurs when part of the mind is already organizing the session into note form—tracking symptoms, interventions, risk, phrasing. Even when notes are written after the session, the cognitive load lingers. If AI systems can reliably reduce this burden, the potential benefit is not only efficiency. It is also availability.

More attention to the client.
Less internal fragmentation.
More capacity to remain with what is unfolding, rather than anticipating how it will later be documented. In this sense, AI may be creating conditions in which empathy is more possible.

This was the last episode of season 3 of the Making Therapy Better podcast. But we would love to hear from you.

Is there someone you would love to see on the show?
Is there a topic you’re eager for us to explore in depth?

If so, send us an email at: makingtherapybetter@carepaths.com.

We would be glad to consider your suggestions for future episodes.

Also, stay tuned as we are preparing exciting new projects for you, including a series featuring Bruce Wampold and Mick Cooper!

Thank you for being part of this community.

Warmly,

Geissy Araújo, Ph.D.

Making Therapy Better Team